Use → or Space to move forward · Esc for the table of contents
Our journey today
One question, 70 years of trying to answer it:
“How do we get a machine to understand words?” Alan Turing was already asking a version of it in 1950: “Can machines think?”
Alan Turing, 1951 Photo: Elliott & Fry · public domain, via Wikimedia Commons
1 · The Secret
What ChatGPT is actually doing (you'll play it as a game)
2 · Words → Numbers
How machines read: from rules to word maps
3 · Learning to Read
RNNs, forgetting, and the attention trick
4 · The Transformer
The 2017 invention that changed everything
5 · GPT & Beyond
ChatGPT, AIs that see, and robots that act
6 · The Big Questions
What AI still gets wrong — and what's next
Chapter 1 · The Secret
You already used AI today — probably before breakfast.
Face unlock
Recognizes YOUR face out of 8 billion
Photo: mikemacmarketing · CC BY 2.0 · Wikimedia Commons
goingtothe
Q W E R T Y U I O P
A S D F G H J K L
Autocomplete
Your keyboard guesses the next word… remember this one.
Your feed
TikTok / Reels learned exactly what makes you scroll
Photo: Solen Feyissa · CC BY-SA 2.0 · Wikimedia Commons
“Hey Siri / Alexa”
Sound → words → intent → an answer, in a second
Photo: Smart Home Perfected · CC BY 2.0 · Wikimedia Commons
But one kind of AI feels different — it talks to you. How?! Let's find out — by playing a game.
Chapter 1 · The Secret
…and it has gone far beyond homework help
Taxis with no driver
Waymo robotaxis carry paying passengers in San Francisco & Phoenix — every day, empty front seat.
Photo: Oleg Yunakov · CC BY-SA 4.0, via Wikimedia Commons
A Nobel Prize
AlphaFold predicted the 3-D shape of ~200 million proteins — its creators won the 2024 Nobel Prize in Chemistry.
AlphaFold structure: SharnKB · CC BY 4.0, via Wikimedia Commons
In the operating room
Surgeons already operate through robotic arms; AI assistance is arriving in hospitals right now.
Photo: Marcy Sanchez, U.S. Army · public domain, via Wikimedia Commons
All different systems — but the loudest revolution is the one that learned to talk. Let's find out how.
Chapter 1 · The Secret
Let's write a story — one word at a time
These are real predictions from GPT-2 (OpenAI's 2019 model, running on this laptop). Click a token to add it to the story!
Once upon a
Congratulations — you just did GPT-2's job. “Once upon a time, the world was…” — it picks from a probability list like this for every single token it writes. Notice the comma was a choice too!
Every percentage here is a real GPT-2 output (computed along the top-choice path). The full menu has ~50,000 tokens — we show the top 4.
Chapter 1 · The Secret
ChatGPT's entire job:
“Given all the words so far, guess the next word.”
That's it. That's the secret.
Do it once — you get autocomplete.
Do it brilliantly, billions of times in a row — you get essays, code, poems, and jokes.
In math-speak: the model learns P(next token | everything so far) — one probability for every token in its vocabulary — and writing is just sampling from that distribution, over and over.
The rest of this talk: how we taught a machine to guess that well.
The map · stop 1 of 5
First stop: turn words into numbers
✓ So far: we know the one trick behind it all — guess the next word, brilliantly. ✗ Still broken: computers can't even SEE words, only numbers. Every idea in AI language starts here: how do we write meaning as numbers?
Chapter 2 · Words → Numbers
Problem: computers don't speak English. They speak numbers.
“hello”
↓
104 · 101 · 108 · 108 · 111
to a computer, text is just codes for letters
But knowing the letter codes ≠ knowing the meaning.
The codes for “cat” and “kitten” look totally unrelated.
The codes for “cat” and “car” look almost identical.
So for ~70 years, the real question has been: how do we turn MEANING into numbers?
Chapter 2 · Words → Numbers · 1950s–1980s
First idea: humans write rules. Meet ELIZA (1966)
A fake “therapist” chatbot. Try it — click a message:
The real thing — an original ELIZA session (1966). Public domain, via Wikimedia Commons
The rule it used:
— click a message first —
No understanding at all — just find keyword → use template.
Rules can't cover jokes, slang, typos, sarcasm… Language has too many exceptions to list.
Chapter 2 · Words → Numbers
The big idea that changed everything:
Old way
Humans write the rules, machine follows them
→
New way: Machine Learning
Show the machine millions of examples — let it figure out the patterns itself
Like learning your native language: nobody handed you a grammar book at age 2. You just heard a LOT of sentences.
Chapter 2 · Words → Numbers
Try #1: just count the words
Treat a sentence like a bag of Scrabble tiles — count what's inside, ignore the order.
“Dog bites man”
becomes
dog ×1bites ×1man ×1
“Man bites dog”
becomes
dog ×1bites ×1man ×1
Same bag… completely different news story.
Word counts are useful (spam filters used this for years) — but order and meaning are lost. Next idea, please.
Chapter 2 · Words → Numbers · 2013: Word2Vec
Try #2: give every word a location on a map
“You shall know a word by the company it keeps.”
Words that show up in similar sentences probably mean similar things:
“I drank a glass of milk”
“I drank a glass of juice”
→ milk and juice keep showing up in the same spots → put them close together on the map
Why is this worth doing? Before, every word was a stranger — knowing about “cat” told you nothing about “kitten”. On a map, whatever the model learns about one word rubs off on its neighbors. Learn one word, get its whole neighborhood for free.
In 2013, Word2Vec let a computer read billions of sentences and place every word on a giant map — automatically.
Under the hood: start every word at a random spot. Read the internet; each time two words appear together, nudge them a little closer (and nudge non-neighbors apart). Repeat billions of times — the map organizes itself. Each position is a vector of 300+ numbers, and “similar meaning” literally means “small angle between vectors” (cosine similarity).
cat & kitten: nearly the same numbers! (real embeddings use hundreds of dimensions, not 2)
Chapter 2 · Words → Numbers
The map of words (hover = neighborhood · click any TWO words = similarity score)
—
This is GPT-2's actual word map: positions = PCA of its real embedding vectors, similarities = real cosine values (computed on this laptop). Real data is delightfully messy — notice “robot” floating between the animals and the machines.
Chapter 2 · Words → Numbers
Wait — you can do math with words?!
king − man + woman ≈ queen
The map learned that the direction from “man” to “king” means royalty. Walk the same direction from “woman”… and you land on “queen”.
We checked, for real: in GPT-2's embeddings the closest word to king − man + woman (after the inputs themselves) is queen (71%), then princess (60%). Bigger embedding models pull the same trick with Paris − France + Japan ≈ Tokyo.
Nobody programmed this. It emerged from reading billions of sentences. Meaning became geometry.
The map · stop 2 of 5
Next stop: read a whole sentence
✓ So far: every word has meaningful coordinates now (embeddings — 2013's gift). ✗ Still broken: meaning lives in ORDER — “dog bites man” ≠ “man bites dog”. We need machines that read in sequence… and they'd better not forget the beginning.
Chapter 3 · Learning to Read · ~1990s–2015
Words have coordinates. Now: how to read a sentence?
Idea: read like a human — left to right, one word at a time, keeping a “memory” as you go. This is the RNN (Recurrent Neural Network).
The whole idea in one line: new memory = f(old memory, current word) — in symbols, ht = f(ht−1, xt). The same little function, applied over and over down the sentence.
The→cat→sat→downmemory updates after every word
It worked!
Early Siri, Google Translate (2016), speech-to-text — all RNNs
But it's slow
Must finish word 7 before starting word 8 — no shortcuts
And it forgets
The memory is small — old words get squeezed out…
Chapter 3 · Learning to Read
Watch an RNN forget
Who was happy? To answer, the model must remember “girl” all the way to the end…
memory of “girl”:
100%
LSTMs (1997) patched this with learnable “keep / forget” gates — a big help, but reading stayed one-word-at-a-time, and very long texts still faded.
Chapter 3 · Learning to Read · 2014: Seq2Seq
Translation had the same problem — but worse
The plan: read the whole English sentence → squeeze it into ONE memory vector → write the Chinese sentence from that.
the whole sentence…
Ilovehotpotwithmyfriendsoncolddays
→
…crushed into this
···
→
…then rebuilt
冷天和朋友吃火锅…
Like reading a whole book, closing it, and rewriting it from a single sticky note. Long sentences? Details get lost. There had to be a better way…
Chapter 3 · Learning to Read · 2015: Attention
The fix: stop memorizing — look back
New rule: while writing each output word, the model may glance back at the whole input and highlight what matters right now.
Like doing a reading-comprehension test: you don't memorize the passage — you look back and highlight the sentence you need.
This trick is called attention. Remember the word — it's about to take over the world.
Translating word #3…
I love hot pot with my friends
writing “火锅” → attention shines on “hot pot”, dims the rest
Chapter 3 · Learning to Read
Attention in action (click a Chinese word)
Thicker line = more attention. Notice 火锅 attends to TWO English words — attention handles that effortlessly. An old RNN had to hope its sticky-note memory kept both.
This one idea (2015) made Google Translate dramatically better almost overnight. (Weights here are schematic — real measured attention arrives two chapters from now.)
The map · stop 3 of 5
Stop 3: a new architecture
✓ So far: attention lets the model look back at everything — forgetting: solved (2015). ✗ Still broken: still reading one word at a time — far too slow to ever learn from the whole internet. Time to throw away the queue entirely.
Chapter 4 · The Transformer · 2017
Then 8 researchers asked a crazy question…
Google, 2017 — the paper
“Attention Is All You Need”
If looking back works so well… why keep the slow one-word-at-a-time reader at all? Throw away the RNN. Keep only attention.
Deleted
Reading in order, word by word
Kept
Every word looks at every other word — all at once
Result
The Transformer — the T in GPT, the engine of ALL modern AI
Fun fact: the authors thought it was a nice translation paper. It became one of the most cited papers in history.
Chapter 4 · The Transformer
The race: reading 8 words
RNN — must wait for each word:
Transformer — everyone at once, everyone sees everyone:
Parallel = perfect for GPUs (graphics chips do thousands of things at once). Now we could train on the ENTIRE internet.
Chapter 4 · The Transformer
“Every word looks at every word” — why does that matter?
Because meaning depends on context. Quick — what does “it” refer to here?
You resolved “it” instantly — attention does exactly this: the word “it” looks around the sentence and locks onto its true meaning. Swap one word at the END, and “it” points somewhere completely different!
Chapter 4 · The mechanism · part 1 of 4
Up close: every word carries three little vectors
Thecatchasedit— our question: what does “it” mean here?
query — the question it asks
“What am I looking for?”
“it” asks: “I'm a stand-in word… WHO do I refer to?”
key — the name tag it wears
“Here's what I offer.”
“cat” wears the tag: “main animal of this sentence, did the chasing”.
value — the package it hands over
“Pick me, and this is what you get.”
“cat”'s package: its actual meaning — furry, animal, the chaser.
Where do these come from? Each one is just the word's coordinates run through a small learned table (a matrix). Three tables → three vectors. Those tables are what training tunes.
Question asked, name tags on, packages ready. Next slide: the matching game.
Chapter 4 · The mechanism · part 2 of 4
The matching game: “it” goes shopping (real GPT-2 numbers)
The
q·k: −1.36
8.1%
cat
q·k: +0.95
81.6%
chased
q·k: −2.03
4.2%
it
q·k: −1.63
6.2%
cat 82%
→
“it” walked in purple. It walks out cat-colored — the word literally absorbed the meaning it paid attention to.
1 · Match. “it”'s question is compared with every word's name tag. “cat” wins with +0.95 — everyone else goes negative. (Dot products can be negative: it just means “bad match”.)
2 · Percentages. Softmax turns the scores into shares of 100%. +0.95 → 81.6%. (That's softmax's entire job.)
3 · Collect the packages. “it” takes 82% of cat's package, plus small slices of the others, and mixes them into its new self.
That's attention — every word softly grabs meaning from the words that matter to it. And every word plays this game simultaneously, not just “it”.
These numbers are real: we ran “The cat chased it” through GPT-2 on this laptop and read out layer 5, head 4 — a head that genuinely exists and genuinely does this job.
Chapter 4 · The mechanism · part 3 of 4
One head is a specialist. So you stack lots of them.
The game you just watched is one attention head. Each head learns to hunt for one kind of connection:
who did what to whom
which adjective goes with which noun
how far apart words are
names ↔ facts about them
opening ↔ closing quotes
…plus dozens nobody predicted
Nobody assigns these jobs — heads specialize on their own during training. Researchers discover what each head does afterwards, like biologists.
Why this beat the RNN: any two words connect in one hop — no more information crawling word-by-word and fading. And it's all matrix math, so GPUs run every head, every floor, at once.
probabilities out ↑
Floor 126 (Llama 3's top floor)
⋮
Floor 3 …
Floor 2 — same, on floor 1's output
Floor 1 — all heads look, then a pattern-spotter cleans up
↑ words go in at the bottom
Each floor re-reads the floor below. Low floors catch grammar; high floors catch plot, logic, tone.
Chapter 4 · The mechanism · part 4 of 4
Now the whole machine, live
The cat sat on the
1 · tokenize
→
2 · embed
→
3 · think
→
4 · score
→
5 · pick
It writes ONE token per trip, then starts over — autoregressive generation. Every ID and percentage on this slide is real GPT-2 output.
Chapter 4 · The mechanism · wrap
What you just watched, officially
Its name: a decoder-only Transformer — these days usually called the “Llama architecture”, after the open model that made it famous around 2024. Llama, Qwen, Mistral, GPT — all this same machine in different sizes.
And training? The same trip, plus a correction: the model reads real sentences, its guess is compared with the actual next word (that is the loss), and every knob gets a tiny nudge — gradient descent, repeated trillions of times. Inference = the trip. Training = the trip + feedback.
Scale of it: every word ChatGPT has ever written to anyone = one full trip. A 500-word answer = 500 trips — each touching hundreds of billions of knobs.
The actual diagram from the 2017 paper — every box in it is something you've now watched happen.
Diagram: Daniel Voigt Godoy (dvgodoy) · CC BY 4.0, via Wikimedia Commons
Chapter 4 · The Transformer · 2018–2024
So they made it bigger. Then bigger. Then BIGGER.
Model size in “parameters” — the adjustable knobs a model tunes while learning. Each gridline is 10×!
The shock: skills nobody programmed just appeared with scale — translation, arithmetic, coding, writing poems… This is called emergence, and it's why everyone suddenly went big. (GPT-4's size is an estimate — OpenAI never confirmed it)
The map · stop 4 of 5
Stop 4: scale it into an LLM
✓ So far: the Transformer (2017) — parallel, hungry, endlessly scalable. ✗ Still broken: it's just an engine. Now: feed it the internet, watch new skills emerge with scale, then fine-tune it into an assistant. This is where ChatGPT is born.
Chapter 5 · GPT & Beyond · 2018
2018: two ways to pre-train — BERT and GPT
BERT (Google) — fill in the blank
“I went to [MASK] and ate ramen.”
Random words are hidden and the model must guess them by reading both sides — a cloze test at internet scale. Superpower: understanding. When you search “bank by the river”, a BERT-style model is what figures out you don't mean money.
How you used it: pre-train once… then for every job — spam filter, review scoring, search ranking — collect labeled examples and fine-tune a separate copy. 100 tasks = 100 specialized models + 100 datasets.
Technical name: masked language model (an encoder).
GPT (OpenAI) — guess the next word
“I went to Ichiran and ate [?]”
Reads strictly left-to-right and only ever guesses what comes next — it never peeks ahead. Sounds like a handicap, but it's exactly what lets the model write.
How you use it: just type. The task rides inside the prompt — “Translate to French: …”. GPT-3 (2020) stunned researchers by doing tasks it was never trained for, from a few examples pasted into the prompt (in-context learning). 100 tasks = 1 model + 100 prompts.
Technical name: autoregressive language model (a decoder).
So why did the GPT branch take over?
Generation turned out to be universal: any task can be phrased as “continue this text” — one model covers them all, no retraining, and the interface is just… language. A conversation is simply one more prompt, which is why the chatbot era is GPT-shaped. (BERT-style models still quietly rank your search results — they just never became chatbots.)
Chapter 5 · GPT & Beyond
OpenAI's recipe: G · P · T
Generative
It creates text — word by word, like our game in Chapter 1
Pre-trained
Before you ever met it, it read a huge chunk of the internet
Transformer
Powered by the attention engine from Chapter 4 — the decoder-only variant, which does exactly one thing: predict the next token
= a Transformer that read the internet and finishes your sentences. Sounds simple. The scale is anything but…
Chapter 5 · GPT & Beyond · Pre-training
Step 1: read (almost) everything
Training = the guess-the-next-word game on the whole internet. Each wrong guess is measured (the loss), and every weight is nudged a tiny step in the direction that reduces it — gradient descent, repeated trillions of times.
~10 trillion
words read — books, Wikipedia, code, websites
70,000+ yrs
how long YOU would need, reading 24/7 nonstop
10,000s
of GPUs running for months
$100 M+
electricity + chips for one big training run
To guess the next word on the WHOLE internet, you're forced to learn grammar, facts, logic, style, even jokes. The guessing game was secretly the ultimate homework.
This is what “AI hardware” looks like: an NVIDIA H100 — ~$30,000 each, and top labs use tens of thousands of them. Photo: 极客湾Geekerwan · CC BY 3.0, via Wikimedia Commons
Chapter 5 · GPT & Beyond
Watch it write — and meet the “creativity dial”
PROMPT
“My cat” — given to raw GPT-2 (124M), which can only continue text
TemperatureT = 0.7 · balanced
OUTPUT
All three are real GPT-2 outputs (same random seed — only T changed). Mechanically: every token's score is divided by T before softmax; small T sharpens, large T flattens. Same model, different dice.
Chapter 5 · GPT & Beyond · 2020: GPT-3
By 2020, GPT-3 was amazing. Also… kind of useless.
It only knew how to CONTINUE text — not to answer you:
YouWhat should I eat for dinner tonight?
GPT-3 (raw)What should I wear tomorrow? What is the meaning of life? Why do we ask ourselves so many questions? In this essay, I will explore the nature of questions themselves…
It saw your question and thought: “ah, a LIST of questions — I'll continue the pattern.”
A brilliant engine with zero manners. One more training step was needed.
Chapter 5 · GPT & Beyond · fine-tuning
Step 2: teach the engine to be an assistant
The fix: fine-tuning
Keep training the same model — but on a new, much smaller dataset: tens of thousands of example conversations, each a question followed by a genuinely helpful answer, written and curated by people.
Why it works
The architecture doesn't change at all — it is still only predicting the next token. But after fine-tuning, the most likely continuation of a question is a helpful answer, because that is what its recent training data looks like.
Under the hood: same objective, same gradient descent — only the data changed.
YouWhat should I eat for dinner tonight?
After fine-tuningHow about a hot pot night? If you want something quick and cheap: fried rice with whatever's in the fridge. Tell me what ingredients you have and I'll suggest a recipe.
This recipe — pretrain on the internet, then fine-tune on conversations — is how GPT became ChatGPT, and it's the same recipe behind Llama, Qwen and the rest.
Chapter 5 · GPT & Beyond · Nov 30, 2022
They put it online for free. Two months later: 100 million users.
Time each product needed to reach 100 million users:
ChatGPT: fastest-growing consumer app in history at the time. Your parents, your teachers, and your group chats all discovered it the same winter.
Chapter 5 · GPT & Beyond · 2023–today
Today's model landscape
GPT-4 / GPT-5
OpenAI — the ChatGPT family
Claude
Anthropic — strong at writing and coding (it helped build these slides)
Gemini
Google — lives inside Search, Android, Docs
Llama
Meta — open weights: anyone can download & run it
Qwen · DeepSeek
Alibaba & DeepSeek — top-tier open models from China
…and yours?
Students today fine-tune small models on a laptop!
Different names, same DNA: Transformer + next-word prediction + fine-tuning. But wait — these models can look at photos now. How does TEXT prediction do that?
The map · stop 5 of 5
Last stop: beyond text
✓ So far: LLMs have mastered reading and writing. ✗ Still broken: the world isn't made of text — it's pictures, sounds and actions. Final stretch: same brain, new senses — eyes, a paintbrush, hands.
Chapter 5 · GPT & Beyond · the variants
The LLM family tree — same brain, new superpowers
LLM — reads & writes text
VLM
+ eyes. Photos are chopped into tokens too.
GPT-4V · Claude · Gemini
Image generators
+ a paintbrush. Your words steer a noise-remover.
DALL·E · Midjourney (the artsy cousin)
VLA
+ hands. Robot movements become tokens.
RT-2 · π0 · humanoids
Agents
+ a to-do list. LLMs that use tools: browser, code, your files.
the 2025-26 frontier
The pattern, one last time: tokenize a new sense → feed the same Transformer. Let's meet the first three up close.
Chapter 5 · GPT & Beyond · Vision-Language Models
Plot twist: images can be words too
A VLM (Vision-Language Model) chops the photo into small patches (a Vision Transformer uses 16×16-pixel squares), embeds each patch exactly like a word token, and feeds them into the same Transformer alongside your text.
Once pictures are tokens, one brain handles both: describe photos, solve hand-written math, read menus in foreign languages, explain memes.
Chapter 5 · GPT & Beyond · Image generation
And AI art? It paints by un-noising
pure noisepicture!
← drag me
Diffusion models (DALL·E, Midjourney, Stable Diffusion) train by watching millions of photos get ruined by static — and learning to reverse it.
To create art: start from pure random static, remove noise step by step, and steer with your text prompt — “a corgi astronaut, oil painting”
“Théâtre D’opéra Spatial” — made with Midjourney, it won 1st prize at a 2022 art fair before judges knew. The US Copyright Office later ruled AI images can't be copyrighted — which is why I can show it to you for free Jason M. Allen / Midjourney, 2022 · public domain, via Wikimedia Commons
Chapter 5 · GPT & Beyond · Vision-Language-Action
Final upgrade: eyes + brain + hands
A VLA (Vision-Language-Action model) outputs movements instead of words. Actions are discretized — “rotate joint 3 by +2°” becomes a token ID — so a robot trajectory is literally a sentence the Transformer can predict.
Third time's the pattern: if you can tokenize it, you can predict it — words, pixels… now actions.
A Unitree robot dog — Google RT-2, π0, Tesla Optimus… the frontier: robots you simply talk to.
Photo: Sgt. Amber Edwards, U.S. Army, 2023 · public domain, via Wikimedia Commons
Chapter 6 · Big Questions
AI makes things up — confidently
YouGive me 2 scientific papers about the sleeping habits of dragons.
AI
Why it happens
It predicts plausible-sounding words — “sounds right” and “IS right” are not the same thing!
It's not lying
Lying needs knowing the truth. It just… autocompletes. This is called hallucination.
Your job
Verify important facts. Great assistant, terrible oracle. Especially for homework
Chapter 6 · Big Questions
The deep one: does it actually understand?
The “Chinese Room” (1980)
A person who knows ZERO Chinese sits in a room with a giant rulebook. Chinese notes slide in — they look up the symbols, follow the rules, slide perfect Chinese replies out.
Outside, it looks fluent. Inside… does anyone understand Chinese?
Philosophers still argue about this.
Team “just statistics”: it's a parrot with a gigantic memory — patterns in, patterns out.
Team “real understanding”: to predict text about the world THIS well, it must have built some internal model of the world. And hey — aren't brains pattern machines too?
There is no agreed answer. You get to make up your own mind — scientists genuinely haven't.
Chapter 6 · Big Questions
Three honest problems
Bias in, bias out
Models learn from OUR internet — including our stereotypes. Ask it to draw “a doctor” and “a nurse” and check who it draws…
Whose art is it?
Trained on millions of artists' work without asking. Artists are suing. Courts worldwide are deciding right now.
Work is changing
Calculators didn't end math careers — they changed them. The winners were people who learned to USE the new tool.
These aren't reasons to fear AI — they're reasons to understand it. Which, congrats, you now do better than 99% of people.
Chapter 6 · Big Questions
Your turn — pick one, argue with your neighbor
1 — If an AI writes your essay and you edit it, whose work is it?
2 — Would you trust an AI doctor that's right 99% of the time — more than a human doctor who's right 95%?
3 — The Chinese Room: fluent from outside, empty inside. Is that also true… of the AI? Could someone claim it about you?
Two minutes, then we compare notes.
Wrap-up
The whole 70-year story, one line
Notice the pattern: every breakthrough fixed the previous one's biggest weakness.
Wrap-up
If you remember only 3 things…
1
ChatGPT = next-word guessing, done astonishingly well
Everything else — essays, code, conversations — is that one trick, repeated.
2
Every breakthrough fixed the last one's weakness
Rules were brittle → learn from data. Counting lost meaning → word maps. RNNs forgot → attention. Attention was slow in RNNs → Transformer. Then: scale.
3
Words, images, actions — it's all tokens now
GPT writes, VLMs see, VLAs act. Same engine, new senses. The story is still being written — possibly by someone in this room.
Wrap-up
Want more? Start here (all free)
3Blue1Brown
“Large Language Models explained briefly” — the most beautiful animations on this topic. Start here.
Andrej Karpathy
“Intro to Large Language Models” (1 hr) — by the legend himself. Then “Let's build GPT” when you're ready to code one!
Jay Alammar
“The Illustrated Transformer” — the classic visual walkthrough, one diagram at a time.
Play right now
ChatGPT / Claude / Gemini are free — ask one to explain today's talk back to you — then grade it.
Build one
Know some Python? Search “nanoGPT” — you can train a tiny GPT on your laptop, on Shakespeare, tonight.
Thank you!
Questions? the entire field started with “can machines think?” and nobody has answered THAT yet either.
Slides: one self-contained HTML file — arrow keys to explore, every demo is clickable. Take it home!
